🆕｜sd3

[noble coyote]

(I'm a beta-tester at Topaz)

[toxic bone]

a lot of people find it to be good nuff. I find diffusion models to be superior with the fine grain details. SDXL and others. Flux showed up and absolutely obliterated the current landscape in that regard.

nice thing about topaz is it's ready to go for a print shop to drop into their tech pipeline. it's not the best way to go about it though, and since we're at the bleeding edge, why rely on legacy solutions?

[craggy crest]

because it's not a legacy solution, and they are fairly bleeding edge - they update their models constantly - and they are industry standard

and most people are not doing anything that requires even that much post production

adobe just put video into firefly btw

[toxic bone]

they're a commercial product. cutting edge is their best hope. Since they don't release their R&D openly, the bleed they have internally is cut and not used by us. Potentially worked inot a future cutting edge product.

It's not diffusion though. it's still a gan. They've really made one that suits studios well, but it's not bleeding edge capable

[craggy crest]

not used by you, but used by a very large number of members of the community, and GANs are actually better in a lot of ways than diffusion models

[toxic bone]

the nature of open source is nightly releases. or pull requests merging in. the "cutting edge" is the stable release

[craggy crest]

and? i don't think the question had anything to do with 'bleeding edge' just 'how can i succesfully upscale flux, i can't get it to work in comfy'

[toxic bone]

i know why they're used. i use a gan to boost the image before i do the diffusion denoise on it. they do well. but ultimately, you zoom in on those highest details, it's recognizeably topaz

[craggy crest]

not everyone wants to be on the bleeding edge - they just want it to work

[toxic bone]

you assume a lot about what i'm saying and where i'm coming from and what i use and how i used it. please stop following me around countering everything i say

[toxic bone]

yeh i covered that one chief

[craggy crest]

i'm not following you around - and you jumped into the conversation not me

so you're countering me

[toxic bone]

#💬｜general-chat message just matching your energy.

i was giving funnyguys advice to listen to that uses "free" tools. not proprietary $200 tools. wasn't any need to argue any of that nonsense and chaos.

i dont know what purpose having a fanboy discussion about a 3 figure priced tool, when you could talk about it's capabilities instead

[craggy crest]

and it's good advice, it's jsut not necessarily going to do him a lot of good - and i did suggest the free image upscaler on capcut's magic tools page as well

[toxic bone]

new features don't really matter at all, when diffusion gets the highst zoom details better

[craggy crest]

well perhaps you could trouble shoot his workflow then, as he isn't successfully getting that

[toxic bone]

i think he got it covered. experimentation is a process that has to be done. i'm a big believer in allowing people to fail until they don't

[craggy crest]

he wouldn't be asking for help after failing enough to get frustrated, if he didn't need help

most people just give up

[toxic bone]

my advice would be to care less about the original composition. It was generated seconds before anywyas. it's not sacred. it doesn't need to be precisely exact. Generally close to it is nice though

[toxic bone]

i mean, those people, if i try to help them they tend to give up too

[craggy crest]

?? i can understand why you might not feel like helping people if they just blow you off and give up anyway

but that seems a little odd

[toxic bone]

if i see someone struggling with a new thing but persisting, i wouldn't say "try doing something else instead". I actually respect the hell out of the struggle

[bitter hearth]

the struggle is real

[toxic bone]

see there you go, acting like you've got me all figured out. I just have a very different way of "helping" people than you might

[craggy crest]

why is it that you have to, so frequently, drop a negative comment and direct it at someone?

[toxic bone]

dont try to understand me. you're jumping to all the wrong conclusions. i'll tell you right now, you should just stop. Sometimes, peopel do need to be told when trying something else is a good idea.

[craggy crest]

i'm not trying to understand you - just have a neutral, non-argumentative conversation with you.

[toxic bone]

i never said i don't like helping people. so, you forced me into this position by saying "i understand that you dont..."

[craggy crest]

or maybe you just took it that way.

[toxic bone]

right. because there was another take away

[craggy crest]

sure there is. there are as many different ways to take a statement as there are human beings on the planet

[craggy crest]

those are really cool

[sacred jewel]

Thanks...

On the large one, it is the exact same prompt, but just 1920x1080 and 50 steps... Sometimes I get lucky straight out of Flux 😉

[toxic bone]

"the offense you took to my words isn't valid" is all i'm hearing. I've heard other people act that way before too. I don't think it's valid. but it's negative to point that out lol.

No one wants this drama. upscalers shouldn't have this much controversy. Diffusion models beat gan approaches in fine grain detail. that's just facts.

i'm out

[sacred jewel]

oh and finally messing around with the max and base shift (as if I know what the ehck I am doing 😛 )

[craggy crest]

might be what you're hearing, isn't what I said. there isn't any drama - you stated some good technical information, we discussed it. you're apparently angry for some odd reason.

[toxic bone]

#💬｜general-chat message you tend to be hostile towards me off the line. these moments persist long past you've forgotten them.

now, i'm out

[dusky thistle]

depends on the scale

and if you've trained a lora for it

i can generate directly at 1920x2560 tbh

no upscale step or tiling

[cinder junco]

With the same level of detail as 1MP? I've tried generating at ~2MP a handful of times and I always got blurry output. I really like 4k+ generations, so tiled upscale works better for me. The Mixture of Diffusers method produces some nice results, but it's about 50% slower than SD Ultimate Upscale, unfortunately.

[sage burrow]

Wow, SDXL sucks beyond beleif!!!!

I was too busy using SD3 and Flux I'd forgotten how utterly horrible SDXL was/is. Fortunately there are hundreds of SDXL loras...

I guess I've been spoiled by Flux!

[cinder junco]

I had a lot of fun with SDXL.

But Flux's level of detail is just 🤌

[sage burrow]

Prob with 20 or so loras.. 😄

[craggy crest]

never learned how to use SDXL?

[sage burrow]

SDXL's anatomy is atrocious! We are talking SD3 ladies on grass level lol

[craggy crest]

yup. never learned how to use sdxl

[sage burrow]

I thought I had, but then today I thought I'd try to create an owl lady riding a horse.....

I've always used loras or checkpoints with SDXL....

[craggy crest]

ah. so you were using base sdxl?

[sage burrow]

A pretty terrible SDXL checkpoint. Well terrible for what I'm trying to do. I'm going to resort to the checkpoints I'm familiar with and talk them into what I need lol

[craggy crest]

here's a starting point

grab the full sized version and the workflow's in it

[sage burrow]

I also tried SXL without any loras or checkpoints, and it reminded me of SD3 laying on grass images

[craggy crest]

well it's a base model...

[sage burrow]

Thank you very much 🙂

[sage burrow]

Why was everyone laughing at SD3 in that case?!

[cinder junco]

Never used a LoRA until I tried one a couple days ago with Flux.

[craggy crest]

i'll DM you

[craggy crest]

speaking of SD3 - here's a good SD3 workflow for you to play with

[sage burrow]

Now SD3 I fortunately haven't had many problems with.

[sage burrow]

You never used SDXL loras?! I have an entire HD full! (but none for women lololol)

Some of the best loras and Checkpoints are SDXL ones 🙂

[craggy crest]

a lora for men works well if you use a female name in the prompt and the pronoun she

[sage burrow]

I'm realizing that instead of just trying to remember which images contain which workflows, that I maybe shold organize them, save them, and put them into folders!!!!!

How do you all organize your workflows so you can find them easily?

[craggy crest]

like that

[sage burrow]

Very detailed too! 🙂

All I need is an owl lady riding a black horse, with a white bg. You can guess how they keep turning out lol

[cursive frigate]

That is some fantastic organization right there.

I should do more of that with my workflows... My images are pretty organized though.

[dusky thistle]

recent research has found that a highly organized lifestyle is associated with an increased risk of dementia

as is picking your nose

[craggy crest]

i have to stay organized, it's too easy to get lost and confused otherwise

[dusky thistle]

confusion helps build new neural pathways and keep your mind fresh as you age

[craggy crest]

recent studies have found the researchers are out of good topics to get research funding for

[craggy crest]

so does reading upside down, writing backwards, and doing other unusual, out of the norm things

[dusky thistle]

however, confusion, and especially frustrated confusion, are especially neuroprotective

[craggy crest]

yes, well - when it turns to irritation and frustration, it's lost it's beneficial effects

(and i get plenty of frustration from other things)

also - always boil your tap water before you filter, then drink it

[craggy crest]

describe for me what an owl lady looks like

[sage burrow]

Not organizing my workflows sorta worked, until we now have 4 models to use 😄 Even just organizing it all by model would help.

[craggy crest]

that's not the description of an owl lady

[dusky thistle]

i've found it's easier when you just mash the keyboard when you save a workflow

the length and content of consonants vs vowels can help indicate the mood and style of the workflow

[sage burrow]

Learning a few model version each day just has to help as well 😄 What was that most recent faster one called? lol

[dusky thistle]

and in place of organizing files, i've found it's better to just shit them all into an external hard drive and keep buying new ones when they fill up

plus, it's neuroprotective

[sage burrow]

is Seagate giving you a kickback for this promtion 😉

[craggy crest]

poke

[sage burrow]

So far Flux is the best at seperating an owl lady and a horse, and giving a white bg...

[dusky thistle]

also, if you save workflows with random file names every time you realize it's an important one

statistically, the odds of finding that important workflow become pretty good

especially if you scatter it across a lot of different folders

[craggy crest]

becky - describe for me what an owl lady looks like

[fleet meteor]

The lighting 😍

[young blade]

made a custom node so i dont have to change the prompt much during lora testing, might add it to the custom node library in the manager at some point if others think it would be helpful

[dusky thistle]

Yeah you need to train a Lora though

[sage hinge]

is SD3 new? better than XL?

[muted dove]

It is newer that XL, but has some "issues".

[noble coyote]

Ollama txt2img Mobius Checkpoint

[limpid thunder]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[sage burrow]

Definitely

[bitter hearth]

flux is really good at the tiny details

you can get good tiny details on SDXL like that but only if you downscale from a larger image

[hallow lion]

Friday has arrived.

[hallow lion]

Reminds me of syd Mead style.

[muted dove]

Oh yes 🙂

Nope, not who I was thinking of 😄

He did Blade Runner and Tron visuals, but it's not like those.

https://www.bing.com/images/search?q=syd+mead&form=HDRSC3&first=1

[noble coyote]

Ollama img2img

[hallow lion]

https://images.fineartamerica.com/images/artworkimages/mediumlarge/3/42-syd-mead-ghazza-lind.jpg

[noble coyote]

SD3@ClipDrop with prompts via desktop Ollama

Breakfast in Brokensville

[noble coyote]

Flux using Ollama prompts

[candid current]

A vector-style illustration of a male and female Polo Ralph Lauren bear playing pickleball, wearing an athletic uniform, vector style

[fleet meteor]

"numbered step-by-step drawing from sketched pencil outline to drawing of an owl on a tree branch"

(flux schnell)

[cursive frigate]

That looks fantastic. If you don't mind I am curious to know what you used for a prompt for that one?

[noble coyote]

A chaotic doodle that forms the shape of a child curled up in a corner. The entire drawing looks like random, messy scribbles, but upon closer inspection, the scribbles come together to reveal the shape of the child and the corner they are in. Around the child, the chaotic thoughts are part of the scribbles, with erratic lines and abstract shapes blending into the figure, symbolizing the mental chaos. There is no pencil effect in this version, making the doodles pure and simple without any textured shading.

...

A dynamic, angled view of an RGB-lit gaming keyboard and mouse, placed on a desk setup. The perspective is diagonal or low, giving a dramatic and futuristic feel, with the lighting from the RGB colors (blues and purples) standing out against a dark background. The keyboard and mouse appear large and prominent, with reflections and shadows creating depth. The surrounding elements like the monitor are subtly visible in the background, adding to the sense of space and modernity. The composition is sleek, minimalist, and optimized for an elite gaming setup, with a cool, high-tech atmosphere.

[noble coyote]

DallE Theme of the Day!

[cursive frigate]

Is there anything available that is capable of extracting prompt data and or tags for an image? Also is there anything that can take an already created image and describe it properly to create a new prompt? Even for NSFW content?

[noble coyote]

Ollama and/or Florence2

[cursive frigate]

Are those custom nodes?

[fleet meteor]

Im not sure if tags but description , the best local model I´ve tested (and run in my gpu ) is minicpm-v:8b-2.6-q8_0 on ollama, but i think florence is better

[bitter hearth]

was gonna say florence

[noble coyote]

Florence2 img2img then save the prompts generated to text file

[cursive frigate]

Do you know if there is a link to a workflow that has that already working?

[noble coyote]

Try my Ollama setup - it can generate separate text files/prompt - or join them all into one massive text file

This is Ollama img2img

[noble coyote]

I updated ComfyUI and my Florence2/Flux workflow broke!

My Ollama w/f does work - it is embedded in the two PNGs above ...

[cursive frigate]

Thank you. Does it matter which SDXL models are used?

[noble coyote]

No, if its only to harvest the prompts.
But you can use SD3, Mobius - if you can get it to work with Flux - let me know! 🙂

[cursive frigate]

If you send me your broken flux workflow I might be able to make some changes and get that working.

[noble coyote]

OK, I'll keep that in mind. Thanks 🙂

[cursive frigate]

What folder do I put that Ollama model in?

[sacred jewel]

Thank you...

I believe it's this one... not sure because I use a large prompt file and randomly pull a prompt. so it could be this one or the next one that was pulled.

(((Ultra HD quality details, realism, ideal))), typography dark moody atmosphere, zentangle In the black void, a white porcelain female face with geometric gold patterns on it, navy blue background. Intricate, abstract, monochromatic, patterned, meditative, highly detailed, dramatic, mysterious, the moody atmosphere of navy blue, the soft misty atmosphere of Russ Mills, the gritty tones of navy blue of Dave McKean's collage work, dreamy and Whimsical magical realistic illustration style navy gray with pink elements of dripping paint - hyperrealism, surrealism. Extreme close-up, poster, fashion, 3D rendering, illustration, typography, dark fantasy, graffiti, painting, film, photo, ```

[hallow lion]

Not only it's Friday. It's the 13th. Well dang.

It's like the final boss of Fridays I guess.

[noble coyote]

Custom nodes

Go to ollama.com and it'll tell you how to setup ollama and d/load some llava models

Ollama img2img

[cursive frigate]

I'm back had to pick up my son from school. Do you recommend a particular model?

[noble coyote]

Llava 3

[cursive frigate]

This one?

[noble coyote]

Yes

[cursive frigate]

kewl... Also I may have a way to get that florence2 workflow working. I found your workflow in an earlier image.

I updated this morning so Ill see if mine is broken too. Once I get it set up.

[noble coyote]

As I said, my Florence2/Flux w/f broke when I recently updated ComfyUI

Hopefully it won't be too long B4 its fixed

Ollama

[shut sable]

.5

[cursive frigate]

Lol I just realized the workflow I found of yours wasn't Florence2 Img2Img. Looks like its Text2Img

[noble coyote]

img2img

[cursive frigate]

Does this Ollama LLM need to run as a separate server or does it just run right inside ComfyUI with no external setup?

[mortal mesa]

seperate app running in backround

Ollama is software that can run many different LLM

[noble coyote]

Initial onetime setup is explained at ollama.com (which is external); and then when using ComfyUI it self-installs

[cursive frigate]

Ok so I got it installed but it doesn't have a UI How do I get a model downloaded into it?

[noble coyote]

At ollama.com go to Models and select whichever model u like and d/load

llama3-8b-text-q6 is good

llava:7b-v1.6-vicuna-q2

[cursive frigate]

so it looks like it does a fresh install every time I want to use a new model?

[noble coyote]

Everytime u d/load a new model it takes a little time. But once d/loaded, it is loaded into ComfyUI instantaneously each time

So just d/load the one Model File for now

Ollama img2img

[cursive frigate]

the one you suggested earlier llava-llama3 Is that uncensored?

[noble coyote]

I don't know ... but you can always try?!

I mean, I have some excellent vanilla (SFW) prompts; yet some unprompted bare breasts do appear from time to time

[cursive frigate]

I ended up going with https://ollama.com/mannix/llama3.1-8b-abliterated

It seems to be working.

@noble coyote Does the Florence2 i2i Workflow work in the same way where it also needs an external instal to run?

[noble coyote]

No, install into Custom_Nodes direct from Github

There is a Florence2 node, and a Florence-2 node - its this last one that I use

Ollama img2img

[cursive frigate]

Whats mobius?

[noble coyote]

A checkpoint https://civitai.com/models/490622/mobius

[bitter hearth]

It will redefine your reality

[cursive frigate]

So with the ollama workflow. It doesn't seem to see the actual image. It still wants a description provided by me and guidance for what I would like the outcome to be. I'm not sure if I am missing something but it just doesn't seem to see the provided image.

[cursive frigate]

Does anyone know which one of these is better for NSFW content?

[brittle nexus]

none

[cursive frigate]

Any Suggestions?

[brittle nexus]

Looks like Joycaption can do nfsw but I never tested

[cursive frigate]

Is that a custom node?

[brittle nexus]

Yes. You can test online too

https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha

[sacred jewel]

dolphin-llama3 does quite well

dolphin-mixtral is another well received uncensored one

[fleet meteor]

Cute 🥺 🐱

[noble coyote]

This is my Ollama "seed phrase"

Ollama img2img (technically img2txt, then txt2img)

Ollama img2img (technically img2txt, then txt2img)

[sage burrow]

Joy caption

[sage burrow]

Joy caption Joy tagger

I've tested both extensively 😉
They don't work very well for sfw though, which is funny

[noble coyote]

Ollama img2img in ComfyUI.

[noble coyote]

Ollama img2img in ComfyUI

[bitter hearth]

@bitter hearth this one

had impact pack workflow, very simple

[noble coyote]

Anybody got an Ollama img2img setup using Flux.Dev at all?

[sacred jewel]

stand by

So sorry... mine is Florence 2 😦

I deleted the custom nodes for Ollama Vision img2img 😦

[noble coyote]

I keep getting some 4 bit/quantisation but need 16 bit/quantisation error (I think?!) 😄

Yes, trying to marry Ollama to Flux gets me 4 is expected but 16 received error (let me find its exact wording)

[sacred jewel]

Tr ythis one

https://github.com/Ling-APE/ComfyUI-All-in-One-FluxDev-Workflow?tab=readme-ov-file

[noble coyote]

Cool!

[noble coyote]

Now that Man Utd have beaten S'oton 3 - 0 - I'm going to have a very nice Saturday afternoon!!! 😄

Dunno what's wrong with Florence2 or TensorOps - cannot get them to load!!!

[sage burrow]

Am I missing something? Civitai just doubled their price for a faster and 1/10th the amount of epochs flux lora training version; then offered it for half price. ?????

In most cases I've only ever found epocks 5 and 10 to be any good, and 1 to be the worst.

[noble coyote]

1st pass Flux/Ollama

[dull star]

left 4 dead 1 lora kind of working wow

[noble coyote]

What is wrong with this Ollama/Flux workflow?

[mortal mesa]

curious what models you using with Ollama cuz there are thousands 🙂

[noble coyote]

Please 😄

Flux.Dev, Flux.Dev.Fp8

[mortal mesa]

what model in Ollama

what LLM or LVM

[noble coyote]

Llava-7b-v1.6-vicuna-q2_K

And Llama3:latest; and Llava:latest

[bitter hearth]

I used mistral 7b for text once

in Ollama

[mortal mesa]

cool just looking for better flavors maybe, atm using Llama3.1-storm.8b, seems significantly better than base llama3.1, just wondering others experiences

[icy drift]

Waiting for this inpainting to be available in Comfy. https://huggingface.co/alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Alpha

Also here's a shiny anime card.

[bitter hearth]

can't tell if that is AI or a photo

[icy drift]

And now the world will never know.

Here's the best 4K I've been able to get. I just can't get rid of that banding / grid pattern no matter what I try. But it's not super pronounced. I officially have this as my new wallpaper until we get a better model.

I'm just floored by its flower accuracy. I asked for:
...a bright covering of variety, including poppies, sunflowers, bluebonnets, daisies, marigolds, and others.

[bitter hearth]

its still very nice

was this with realism lora

[icy drift]

No, but it was a finetune from Civitai plus the Hyper-SD lora. Generated in 3 passes of 6 steps each, around 1 minute on my 4090 if I remember right.

[bitter hearth]

wow you can use just 6 steps?

this is with 8 step hyper lora?

[icy drift]

Pretty bad results at just 6 steps. But I gen at 6 steps, sharpen it, then denoise 0.63 or so at another 6 steps, upscale with a model and sharpen again, then do a final denoise at 0.41 to 0.57 for 6 steps. So 18 steps total. The sharpening helps add detail like texture.

[bitter hearth]

ah nice that sounds good
yeah I use that smart sharpen from Comfy Essentials, sharpening between passes helps a lot

[icy drift]

I'm re-running the workflow now. I was doing an initial 1920x1080 render, then upscaling it to 3840x2160. It's taking 159 seconds per image, so about 2 and a half minutes.

For comparison, here's my previous desktop from Cascade, which was about a 30 second 4k native render if I remember right. The overall impression is great, but the reflections are wrong, and the flowers are just sort of nothing weirdness.

[bitter hearth]

ah yeah I haven't got into cascade yet

[icy drift]

It doesn't have prompt accuracy, and it can't do details, but it can do very large images. I haven't used it in a while now. Just Flux.

Here was my SDXL wallpaper before we had Cascade. Definitely no detail, and the reflections are subtly off for the mountains, and either missing of off for the near shore.

But each of these was the absolute best 4K I could get out of a given model after a huge amount of experimentation and cherry picking.

[bitter hearth]

that's rly nice yeah

these two are my best high res images, using SDXL
they have comfy workflow attached if you want it

method was resadapter, gradual deep shrink and perturbed attention guidance

[icy drift]

Yeah those nothing-details work really well in a jungle setting. They could be vines or leaves or moss. The panels on the robot don't really work though.

[bitter hearth]

yeah its a bit messed up

if I remember rightly it had FreeU node with high S values
which makes more fine details but messes it up a bit

that effect where the panels are a bit wobbly

[icy drift]

I got a lot of missing red nodes. But SDXL always messes up R2's panels. Flux was the first model I ever saw make them solid (if not exactly movie-accurate).

[bitter hearth]

yeah flux does good R2
there's kinda 2 flavours, clean and dirty

[hexed dirge]

https://civitai.com/models/752718

[noble coyote]

... An epic and symbolic portrayal of Laplace's Demon, standing triumphantly as the Vanquisher of Chaos. The demon is depicted as a wise and powerful figure, exuding an aura of omniscience and control over the universe. In one hand, it holds a scroll or a celestial map, symbolizing its ability to predict the future by knowing all forces and particles. Surrounding the demon is a swirling vortex of chaotic elements like flashes of lightning and fragmented celestial bodies—being subdued and ordered under its will. The demon itself is calm and composed, with glowing eyes and an ethereal presence, while the chaotic elements appear tamed in the background. The overall scene evokes a sense of control over the unpredictable, combining cosmic imagery with an air of philosophical depth.

... A tilt-shift effect scene featuring 3D cartoon people in a room, running around in a state of panic. The characters are exaggerated with comical expressions and dynamic poses, emphasizing the chaos. The room is filled with various office items scattered around, adding to the frantic atmosphere. In the center, a large speech bubble with bold, capitalized text reads 'STARBOARD IS DOWN'. The overall style is playful and cartoonish, with vibrant colors and a miniature, diorama-like appearance due to the tilt-shift effect.

... "Abstract impressionist painting with heavy impasto paintstrokes and thick textures. Dark cosmic colors. A painting vaguely depicting the Earth, hovering in a pitch black void. Accenting faint color streaks and ribbons. Undifferentiated shapes and indistinguishable matter. Faint blurs and digital grunge effect."

Help?! Why is this so slow?! #🆕｜sd3 message

[mellow edge]

How good IS SD3 for logo?

[noble coyote]

DallE Theme of the Day

[sage burrow]

Via Mage (so schnell)

[sage burrow]

They should have hired me to make the images 😉

[pseudo owl]

it's decent but unless you are using the 8b api, I would probably recommend flux instead.

[sage burrow]

Schnell doesn't believe in male fairies it seems

[mellow edge]

Thank you. It's possible for you to give me some samples?

[pseudo owl]

for sd3 or flux?

[mellow edge]

With flux if you Can

[mellow edge]

Something in the the gaming niche if possible

[sage burrow]

Schnell via mage

Well OK not always rofl

[mellow edge]

Thank you. Looks better than SD3

[sage burrow]

This us just scnell, dev and pro ate even better!

[mellow edge]

I like it

I have tested with SD3. But its never comes with great résults

It's dont write all the letters

Maybe a Bad prompt from myself

[sage burrow]

You have to reroll a lot sometimes

[sage burrow]

Are u using flux?

[mellow edge]

Not yet. I use SD3 with api as i dont have a pc with GPU

[sage burrow]

Took 6 tries. That's schnell tho, dev is better

[sacred jewel]

Vintage X-Ray Machine?

[sacred jewel]

https://civitai.com/articles/7379

[bitter hearth]

wow this is an amazing study
seems like small differences affect it a lot
couple of things I noticed is I liked higher step ones more, and the background trees mostly only came back with high guidance

[sacred jewel]

Yeah, I also noticed in some cases, details come and go depending on step count.

[bitter hearth]

its very strange
I did a similar study with shift, sadly I forget to save stuff
it would alternate between good and bad image sometimes

[sacred jewel]

Yeah, that is likely my next test is the whole shift thing.

I long for the day that we just ... wait for it....

https://tenor.com/view/ron-popeil-set-it-and-forget-it-set-it-forget-it-showtime-rotisserie-gif-19856802

[bitter hearth]

haha yeah

we're still pretty far from that

I did a test of 1000 steps once:

definitely not worth it for most images

if I remember rightly this is guidance 1.2-1.4 and shift 0-0.3

[sacred jewel]

Cutaway LoRA

[noble coyote]

Flux/Ollama - img2txt (Ollama) - txt2img (Flux)

[noble coyote]

SD3 originals Ollama/Flux img2img

[icy drift]

Ooo. It definitely has credit cards in the training data. I wonder, does that mean it has the potential to leak credit card numbers? 🤔 (I asked for the name John Smith and the number 1234 5678 9123, which I realize in retrospect was not a long enough number to ask for.)

This things text accuracy still amazes me though. These aren't cherry picked. I've only generated two credit card images. (Gonna stop though. Nobody likes thinking about credit cards.)

[icy drift]

nlo.com redirects to telepathy.com.

[summer kelp]

Which model is this? I tried to make some wheels with juggernaut XL but it cant make the lug nuts for some reason

[noble coyote]

Pixel art style illustration inspired by a Street Fighter 2 screen. A teddy bear character, designed with exaggerated features like oversized button eyes and a stitched mouth, is performing a dynamic spinning move resembling Zangief's Typhoon. The teddy is wearing a colorful wrestling costume with a cape. Opposite, a human character in a vibrant blue judo gi is standing with legs apart, one arm extended as if launching a hadouken-like projectile, with a glowing hand effect. The background is a cozy sofa room with pixelated details like a sofa, a coffee table, a lamp, and plush toys as spectators. The scene includes pixel art health bars, scores, and a timer at the top to resemble a classic fighting game screen.

[sullen moss]

I'm working on an art piece right now and ran into a problem—FLUX doesn’t understand the skin of older people

Promt: A close-up image of an elderly grandmother’s hands, showcasing the detailed texture of her wrinkled, aged skin. The hands are weathered with deep lines, visible veins, and worn knuckles, symbolizing years of hard work and life experience. The lighting is soft and natural, highlighting the creases and folds, bringing out the warmth and wisdom in the hands. The background is blurred, ensuring the focus remains entirely on the hands, capturing the essence of age, history, and quiet strength.

Flux

Dalle

[errant dust]

Nor does Dall-E per that image

It looks nothing like an old person's hands

[short thicket]

Working on an algorithm for Mangled Merge Flux. I'm doing things old school until I or someone else can figure out the more advanced Dare/Ties/Della methods. Only 20 loras in so far, but seeing some promising results. Still a little rough around the edges though.

Dev on the left and Mangled Merge on the right.

[cursive frigate]

Does anyone know what could cause this at the edges of an image? Also have an idea for a possible solution?

[neon sail]

怎么用

Cutaway LoRA

[short thicket]

40 loras merged. Some outputs seem better, some seem worse. Needs more experimentation. Dev on the left, Mangled Merge on the right.

[craggy crest]

those are printers marks - odd that you'd be getting that on a generated image

[cursive frigate]

I read it might possibly be because the original image is not divisible by 16. This only happens with img2img by the way. but ya, I thought it was a bit wierd and interesting.

[toxic bone]

flux dev tends to do that for me, maybe 2-5% of the time

might be a sampler thing. nature of the mmdit network and the edge of math for dpm2++ . iirc it happens more with that sampler

[noble coyote]

Revealed in a-Dream: SD3.5 has 8 billion paramours!!! 💖 🥳

[gilded silo]

i can try other prompt if you wish to test the model

[open marsh]

howdy yall, so i have been importing specific tools from custom nodes and ive been getting "name 'WHATEVER' is not undefined". are the definitions not in the same .py documents where i get the codes to copy and paste from? o.O

[dusky thistle]

Here is the image you requested.

[open marsh]

o.O i requested an image? also, thanks for the help, but i cant read what that language is in after the title 😅

[noble coyote]

Ollama/Flux img2txt - txt2img. Ollama input = What is this an image of? Make an image which adds a face, head, hands and arms.
[The input images are vintage fashion mannequins.]

Some output remained headless; so I utilised Generative Fill in Photoshop to produce head and hair.

[noble coyote]

The good thing about Ollama is that it is img2txt first - so you can modify the text prompt on the fly - which then becomes txt2img.
With Florence2, editing the prompt is only possible by stopping the process; modifying the prompt generated in the text file; and starting a new txt2img implementation of Florence2.

[cursive sandal]

A group of wild boars and a group of hyenas are confronting each other from left to right. During the day, the atmosphere is tense and panoramic ,in the style of Pixar, in the style of Disney --ar 16:9 - @sturdy tiger (fast)

[sacred jewel]

in how many weeks?

[sacred jewel]

Pink ballz

[sullen moss]

Someone call the moderators, or soon the inappropriate proposals will start. LOL

[toxic bone]

mod's shouldn't be needed for that one. its the same post in every single channel here, verbatim, from a user that just joined.

The server is wide open to abuse when that can happen. it's only a matter of time before all the other ants find out that this place doesn't have mod controls during certain parts of the day

[strong bramble]

#imagine car

Imagine A muscular man stands confidently, covered in dust and sweat, wearing a fitted, worn-out t-shirt that highlights his toned arms. He holds an object casually in his right hand, with a rugged military-style jeep parked behind him in a dusty, outdoor setting. The camera angle is from a low perspective, looking slightly upwards, which emphasizes his imposing figure and strong posture. His hair is styled in medium curls, giving a rugged and natural look that complements his intense demeanor. In the background, a hazy mountain range stretches into the distance, adding to the scene's rugged atmosphere. The lighting accentuates his determined expression as he gazes off to the side, creating an intense and gritty vibe.

[summer kelp]

There is no bot that will do this for you...

[strong bramble]

Y

[summer kelp]

No idea, just run your own ai

[bitter hearth]

there is a bot, if you want one, in the artisan channel

[sterile pendant]

In comfyui, if you set flux guidance to 1.0 and then use a regular cfg of like 4 or 5, it really does allow a negative prompt and doesn't destroy the image. Saw some example on diffusers for "true cfg" and looked through the code and it's easily doable in comfy doing what I just said. Obviously, it doubles inference times though.

My next experiment is going to be doing some hybrid approach with something like dynamic/automatic cfg where it starts out with guidance 1.0 and CFG 5.0 and then transitions back to something like 3.5 guidance and 1.0 cfg. Might need to make a custom node to handle it though

[bitter hearth]

I just saw the reddit post

not sure how they pulled it off

I wish they wrote an explanation lol

[sterile pendant]

literally do what i said, just set flux guidance to 1.0 and cfg to like 5

in comfy. probably works in forge as well

those two images are the same seed btw, 30 steps, dev

[bitter hearth]

wow that's it? such a simple fix all this time

[sterile pendant]

yeah...

it's f'in weirding me out how simple it is...

[bitter hearth]

rly weird how simple it is yeah

cos those other methods mostly have side effects, dynamic thresholding, autoCFG etc

I use tonemapwithrescaleCFG personally, on every image

[sterile pendant]

right, but they probably aren't having people set guidance to 1.0. but that's why i was saying i want to try some kind of transitional approach like what dynamic/automatic cfg do, but have it go from guidance 1.0 cfg 5.0 -> guidance 3.5 cfg 1.0

figuring out the ideal falloff curves might be a bitch though

i know people were hackily doing the perpneg stuff a while back, but it had drawbacks probably related to guidance as well

but f me, flux can actually do art this way now...

[bitter hearth]

perpneg is nice on SDXL sometimes, it acts a bit differently

interestingly perpneg combines with CLIPNegPip (you can weight a token negative)

so if you use perpneg but the tokens in the negative have negative weight, you get a rly strong positive effect

[sterile pendant]

i'm too lazy for all that diminishing returns gains type stuff, i just use automatic cfg or dynamic cfg lol

[bitter hearth]

automatic cfg wins for sheer laziness yeah its great

the reason I have to hand tweak it is I like to use a ton of PAG, SAG and SEG to different blocks
then a really heavy anti-burn setup is needed to counter all of that

this is for SDXL though, flux doesn't have attention guidance nodes yet

[sterile pendant]

automatic cfg can do all that with the advanced profiles, iirc

[bitter hearth]

oh yeah I saw that in the repo but I never learnt how

would probably be ideal for me

[sterile pendant]

like the "excellent attention" preset does PAG

on top of the automatic stuff and boost

[bitter hearth]

yeah I need to learn that part of the repo

at least for my images, PAG, SAG and SEG have been the biggest quality boost in the last year

[sterile pendant]

here's another example of it working, gawt dayum

wow i even forgot to change an to a when i changed the prompt, oops

[bitter hearth]

wow yeah it really works

it does have a couple of downsides to manage, because more guidance generally gave better compositions

and secondly because higher resolutions sometimes needed more guidance

maybe scheduling helps

[sterile pendant]

this is an absolutely basic workflow, using simple and euler

just loaders, modelsamplingflux, cliptextencodeflux and ksamplers

about to try it with my hyper merge 16 steps. i also want to see how 1:1 cfg vs guidance really are. it might be that guidance 3.5 is actually closer to something like cfg 7 or something

[bitter hearth]

I always forget the name but a paper found the best steps to apply negatives
and it was less than the whole steps

the hyper lora looks good yeah

[sterile pendant]

yeah which is what boost mode on automatic cfg does

[bitter hearth]

for SDXL I actually use TCD and two different hyper loras now lol

4-8 steps are great

[sterile pendant]

negatives really only matter early on, basically, where the largest area under the sigma curve is before it goes into "refining details" steps

where the curve falls off

[bitter hearth]

yeah the composition kinda gets locked in

its quite funny in the beta sampler paper
they made charts of what steps different sized details appear in
and they found that almost nothing happens in the middle steps

there's this weird gap

[sterile pendant]

yep, mostly just broad color shifts. i use beta a lot

[bitter hearth]

yeah for SDXL I always use beta now

[sterile pendant]

same

[bitter hearth]

the only exception is when the composition is having trouble

then I try karras, and then SGM uniform

this mostly happens trying to stretch the resolution too much

[cursive frigate]

I know this may seem like a truly stupid question, but, how do you know which one should be Clip L weights, and which one should be T5 weights. I only have one Clip on this custom workflow. With no negative prompt at all. Can I set them both to T5?

Also I have a 4090 and cannot for the life of me get the fp16 weights to work. This stuff make my head spin.

[sterile pendant]

just tested with automatic cfg and dynamicthreshold, it works with both. but with automaticcfg, you have to disable boost or it turns into a mess

[sterile pendant]

here's another example. note how it broke in the middle, but when using automatic cfg, it fixes it on the right. definitely sticking to using automatic with things(boost off because it breaks things)

[gilded silo]

whenever lykon says so

[noble coyote]

FWIW 16 weights work on my 2070

[craggy crest]

more than -1 and less than 10,000

[craggy crest]

yeah, when you turn off flux guidence, you basically turn off flux

[floral gate]

What would you call this style?

Vintage illustration doesn't get me this look exactly

[bitter hearth]

Vintage illustration would have been my best guess too

[frozen oxide]

A small, fluffy cat with white and gray fur is sitting on the surface of the moon, gently nibbling on a fish. The fish is glowing slightly, with a silvery-blue hue, as it floats in the moon’s low gravity. In the background, Earth is visible, casting a soft blue glow across the moon’s cratered landscape. The sky is pitch-black, dotted with bright stars, adding to the surreal and tranquil scene."

[noble coyote]

naive minimalist?

[floral gate]

ohhh that sounds nice

Anyways, I figured it out

For some reason my schenll only liked when I used flat colors in the prompt as well as vintage illustration

Kind of weird but I don't care

[pseudo owl]

This is pretty nice, they made flux used normal guidance scale. Results seems pretty good: https://huggingface.co/nyanko7/flux-dev-de-distill

[frozen oxide]

How i can generate image from text

[floral gate]

By chance, does anyone here use lm studio because I need some help figuring out how to make it use my GPU and not CPU? Other programs don't have a problem with it.

[pseudo owl]

Idk much about lm studio but I know a decent amount about of what it uses(llama.cpp) you just have to set n gpu layers to -1 or like 99

[floral gate]

Can you DM me and then like put that into dummy words

[noble coyote]

https://www.gpu-mart.com/blog/run-llms-with-lm-studio#:~:text=If you would like to,settings for running local LLMs.

[floral gate]

I tried looking but it's outdated or something. There's no section like that, actually, the whole UI Is just different

[noble coyote]

It does the same thing as Jan.ai

[floral gate]

Yeah that's what I'm coming from

I didn't like that Jan forced gguf so I switched to this, but I don't see any gpu acceleration option anywhere

Unlike Jan where it let me and everything was always lightning quick, this takes forever to load a simple sentence off of my cpu

There's one thing in all of settings talking about gpu but it doesn't do anything because lm studio isnt picking up my gpu or its broken?

Like I have the whole cuda stuff updated and I was using Jan a few hours ago

[noble coyote]

Do u have nVidia GPU?

[floral gate]

3060 8GB, which is supported by lm

[noble coyote]

Use GeForceExperience Console to pair LM Studio with GPU

[floral gate]

How do I get there?

[gilded silo]

have they fixed the questionnaire yet?

[noble coyote]

https://www.nvidia.com/en-gb/geforce/geforce-experience/

[floral gate]

Well here's the thing. I've found some weird part of the settings that had like something to do with CUDA and llama.cpp, so I updated that and the cpu, and then also in settings for the model I turned off keep memory in case that works but I don't think it does. And so basically, the program recognizes my gpu so I don't think I have to mess around with nvidia settings as this is clearly an app issue. Also my gpu drivers are updated

But still wont use my gpu at all

[noble coyote]

Jan.ai lets me select GPU

[floral gate]

Yeah, which is why I'm debating going back, but I love the easy-to-use model discover page and also that I'm pretty sure I'm not limited to gguf like on Jan. But if they just removed the gpu acceleration settings then I definitely will

[sacred jewel]

What has your testing shown? They only show 2 examples 😦

[sacred jewel]

https://discord.com/channels/1002292111942635562/1237461679286128730

[toxic bone]

you got nvidia card? cuda needs it. zluda is dead now too.

[floral gate]

Oh shouldve deleted it. I found help and got the issue fixsed, thanks though

[toxic bone]

cool. also, koboldcpp was better imo, but llm studio do work well in one small package

[craggy crest]

it just rambles on about wanting to know the name of your first born son, and what you had for breakfast today

[gilded silo]

i see

[toxic bone]

wild link ty for sharing

i'm going to wait for there to be a version that's separated files. i dont' want dozens of t5 copies on my drive. but it is on my list of things to test and explore this week 👍

[pseudo owl]

It’s only fp16 version of the dit model, no t5, clips or vaes I believe. Sadly can’t fit it in my gpu.

[toxic bone]

BFL published Flux.1 Dev in fp16. The 23gb file has t5 and clip l inside of it.

[toxic bone]

The stock weights when pruned are 16gb. every 23gb version is over 7gb of duplicated data

[bitter hearth]

was so confusing finding which civit version was what

lots of different combinations on there

[toxic bone]

different quants confuse the decision surely

[bitter hearth]

the various FP8 speedups are what I had the most trouble with
turns out there's two versions of FP8 and some of them need a specific version

fp8 e4m3fn vs fp8 e5m2

[toxic bone]

||mfw i got away with calling you Shirley||

[bitter hearth]

Airplane is good

[toxic bone]

i thought that em5 was determined not suitable for flux puproses at all

something about the way the data model aligns

[bitter hearth]

maybe, the comfy FP8 speedups require the fp8 e4m3fn one

[toxic bone]

leave it to people to try it anyways

https://tenor.com/view/it-crowd-bastards-what-a-bunch-of-bastards-people-sip-gif-17774239

[bitter hearth]

I've switched to this pytorch script which works differently to both fp8 e4m3fn and fp8 e5m2
it quants different parts of the model differently https://github.com/aredden/flux-fp8-api

[toxic bone]

neat looking

[bitter hearth]

as far as I know this is the fastest script, it matches Fal.AI's speed if used on H100

flux is way too slow I want to find all the speedups lool

sadly I don't like the hyper loras, that might have been good. hopefully lightning or TCD is coming

[pseudo owl]

try this: https://huggingface.co/DarkMoonDragon/TurboRender-flux-dev
seems like a pretty nice alternative(supposedly more detail and requires 4-6 steps and even better detail at 10-15)

[bitter hearth]

thanks a lot that seems promising

[icy drift]

Inpainting Controlnet with Flux!!! 🎉 🔥

[icy drift]

Workflow. ComfyUI now supports the inpainting controlnet from AliMama https://huggingface.co/alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Alpha/discussions/1

[icy drift]

This was just a proof-test. Generated at 6 steps for the initial render and the inpaint.

[toxic bone]

i dont think seams like that show proof of concept. i guess its a workflow that executes to the end though

[bitter hearth]

don't know if it will load on discord
but this is the img-to-vid for Cogvideo that came out today

[icy drift]

It's proof the controlnet is supported (and compatible with finetunes, with Differential Diffusion, and with the Hyper-SD 8-step lora), not proof of quality.

I am trying with less harsh settings now.

[bitter hearth]

oh no

[pseudo owl]

yeah cogvideo img2vid is nice, this is a pretty good video it made. (original img was low res)

[bitter hearth]

that worked well yeah

got to do the cropping outside really cos it squishes a bit. It pulls of simple pans nicely

[icy drift]

It can do seamless inpainting for sure, but I am getting very unstable results. I don't understand inpainting nodes in Comfy I think.

[pseudo owl]

yeah nice, its taking forever though since there are 10 people waiting in the queue.

[pseudo owl]

cogvideox is really amazing, definitely best open source model in image2video and I tried it with an actual good image(made by flux) and it honestly blew my mind.

[hexed dirge]

quality speaking seems very good, but movement are so unnatural

[pseudo owl]

wdym by unnatural?

[hexed dirge]

she's too steady

[pseudo owl]

oh yeah thats true, but i didn't really specify what she should do in the prompt so its understandable

[bitter hearth]

its a big step up yeah

working out what sort of video works best with it is an important step i think

and definitely the dimensions cos it squishes 1024x1024

[pseudo owl]

i believe it said 720*480 works best, its nice that flux can generate at that res.

[bitter hearth]

yeah its not a bad res

its quite good timing TBH that we got flux for this

given that flux is more flexible with res

[cursive frigate]

Here is a couple of Demonic looking Warriors I created tonight. Some img2img with AI captions then took the caption and put it through text to image with some LoRAs to get these.

[noble coyote]

A prompt constructed through Jan.ai, Flux1-dev -Q8_0

[icy drift]

Did you get cogvideox running on windows? Compiling DeepSpeed requires Intel's OneCCL, which is closed-source and does not have a windows binary available as far as I can tell.
Edit: Nevermind manually installed older version of deepspeed with windows support. Let's hope that resolves everything... Requirements install is progressing!
And finished 😮 Let's see if we can run inference.
Weights downloading now...

[icy drift]

Reinstalled Torch and running now. 😮

[icy drift]

Okay, no open-source image-to-video yet. I kept failing to make i2v work, then finally found this note "(Coming Soon)" next to the 5b-I2V model link.

Image to video from the (currently) closed-weights HF demo.

[alpine summit]

Flux

[noble coyote]

Flux.Dev+LoRAs

LoRAs = Toy Camera, Wraith_BW, Blue Future

[lunar sentinel]

how to use

[gilded silo]

3.5L

[dull star]

SD 3.5?

[cold kite]

beach

[gilded silo]

sd3.5 large

[dull star]

thanks

so 3.5 is a thing

we'll see how good medium is

[gilded silo]

unfortunately, 8b parameters worth wasted on this

all images posted are pre-finetuning (whether it be safety or aesthetic)

[pseudo owl]

Not that bad, try some humans(something like 3 people holding a sign laying in grass, schnell could do it). Doesn't beat flux but could give solid competition, not sure what your prompt is but this is schnell 15 steps, not even dev.

prompt: A photograph of a complicated notebook filled with formulas and a post it note stuck on it that says "there are so many latents left to diffuse" written out in black ink and "Diffusion Sells", high quality, 8k, real, incredible detail.

[dull star]

the parameter advantage really shows with the clarity of the text

and possibly its also more trained

[gilded silo]

80 steps on dpmpp_2m is the sd3.5L default

woman lying in grass

cannot test multi-subject composition, please tell lykon to start the bot again

[dull star]

large didn't suffer from this iirc

only medium

the woman lying in grass

[gilded silo]

safety tuning does that to you

[turbid grotto]

What is this, do you have early access?

[gilded silo]

a sort of early access, yes

[turbid grotto]

any info about status?

[gilded silo]

extremely early access (started 1w ago) points in a good direction, unless they start the aesthetic tuning now

[turbid grotto]

oh, so needs more cooking, nice

[sacred jewel]

https://tenor.com/view/john-krasinski-krasinski-john-wait-what-gif-1509324842126244389

[gilded silo]

always need more t ime in the gpus, never enough time

[turbid grotto]

I think it will be a big deal, if they release training script and don't plan to distill or something else

[gilded silo]

more tim e

[turbid grotto]

Yea, thanks for sharing!

[pseudo owl]

I like the aesthetic, flux(even schnell) seems to be better but its not done training so it could be a great alternative and its not distilled.

[gilded silo]

8b parameters yet lose to a GAN-wannabe

[pseudo owl]

I wish they release 8b instead of 2b this time, people could just quantization to make it use a small amount of vram.

[bitter hearth]

wow so SD 3.5 is real

that grass picture is workable, its good enough to refine with realvis

[turbid grotto]

aesthetic seems to be more pleasant than flux, for my opinion

[bitter hearth]

yeah this is more realistic than base flux

some flux fine tunes are more realistic but not the base

[turbid grotto]

looks very great for pre-aesthetics tuning btw...

or I have wrong understanding how should it look

[turbid grotto]

hope it will also share sd3m's ability to generate unique faces

[gilded silo]

one must hope lykon retains the style through finetuning

[turbid grotto]

yea!

[cunning lintel]

imho the less aesthetics tunig the better, aesthetics tuning gets you flux, push out all aesthetics but the one you want, i'm not a fan

[bitter hearth]

aesthetics tuning doesn't mean not realism neccesarily

[cunning lintel]

but 🎉 SAI/SD3 shows a sigh of life again!

[pseudo owl]

yeah great, the style is nice but it duplicated left and forgot diffusion sells(assuming you are using my prompt)

sd3.5 large is showing promise

[turbid grotto]

They probably have to make comparable base model, but would be cool to get pre tuning version too

[runic tusk]

If you want to nitpick, your image isn't of a "complicated notebook", it's of a textbook with printed text. A notebook would be handwritten notes, so it's not like your prompt was followed any better either.

[turbid grotto]

sd3.5L's generation at 621x621 btw
or it got downscaled somehow...

[gilded silo]

downscaled, native is 1024x

[cunning lintel]

My hope for SD3 is to be reasonable at prompts but look much less AI/plastics/smoothed than flux

[pseudo owl]

did you generate it with sd3.5 large? and yeah the notebook part could be considered wrong.

[bitter hearth]

also to have negatives without tricks

[turbid grotto]

and it should be a bit faster!

[cunning lintel]

when 8b was offered in api i always used to comment about how difficult it was to get different styles, but compared to flux, it was brilliant in hindsight :p

[bitter hearth]

the 8B version of SD3 was always good I think
I don't use API models but if that model was open sourced I would switch to it

2B was where the problems were

[bitter hearth]

these landscape ones are all great especially the shark

[cunning lintel]

I never minded it that much, seeing it as teaser (and gliff offered lots of gens for free), but when the "real" thing turned out 2b and the teased thing never was released, i lost interest. I hope this time there will be weights, after all this time, only actually seeing weights is believing for me

[bitter hearth]

yeah need the weights for sure

they should have just dropped the SD3 8B weights as soon as the controversy began TBH

and then they could monetise SD 3.5 Pro or something

but its easy to forget the company nearly went under this year

[pseudo owl]

Flux with normal cfg produces much more nicer imgs imo, you need good negative prompts tho

[bitter hearth]

oh that's good for flux

what was the negative?

[pseudo owl]

I don’t have the exact one now but it was something like: “distorted, deformed, low quality, average quality, bad quality, night, ugly,”

Prompt was: “a woman laying in grass, sunny”

[bitter hearth]

ok thanks

I don't need it exactly yeah

[fleet meteor]

Aaaa where is iiiiiit 👀

[sterile pendant]

@pseudo owl @bitter hearth another finding I forgot to mention yesterday when I was talking about using negatives with flux is to set the flux guidance to 1.1 and then the cfg to whatever. Kind of like when using cascade where the one stage uses 1.1. My guess is it helps avoid holes or poles(zeros or infs) or plain identities in calculations somewhere, but I'd have to analyze the code more. All I know is that tiny difference helped, but it still seemed to work well into the 1.5 range. Keep in mind, I was using automaticcfg(boost off) with things

[gilded silo]

ask lykon

[craggy crest]

never more!

[noble coyote]

Flux+LoRAs

[sullen moss]

I can't quite understand. SD 3.5 has already been released, but it's again limited by some bot? What's the point of such promotion?

[bitter hearth]

they have to have terrible marketing, its traditional at this point

[lunar rivet]

it might as well not exist if you can't download it from huggingface

[bitter hearth]

yeah but its nice to have previews like this to know its coming

for budgeting purposes

[alpine summit]

Sd3

[sacred jewel]

ballz

[frail shoal]

Very pretty, what Lora is this?

[frail shoal]

It has been released? Where ? How can we try it ?

[bitter hearth]

some people got early access

[errant dust]

https://www.msn.com/en-us/news/technology/new-research-could-make-weird-ai-images-a-thing-of-the-past/ar-AA1qBC6K

[errant dust]

WHere was this reported?

I have a feeling we are only going to get a 'fixed' SD3 Medium, but here's to hoping I'm wrong

[bitter hearth]

doubtful, I think they are gonna release the 8b model, if not then they lose to flux

[errant dust]

I have no such expectations. Still, what I'd really like to know is how reliable that claim of 3.5 is.

[bitter hearth]

they're obviously gonna release something, june 12 was their last release and it flopped so 3.5 in a month or so doesn't seem unlikely, just curious if it's open weights